Scott Hoang comments

Results 21 comments of


                                            Scott Hoang

fasterRCNN compilation error with python 3.8 and pytorch 1.12 cuda 11.3

hi zee-fee, you can download the fix here https://github.com/jwyang/faster-rcnn.pytorch/pull/894

this repo is good ONLY for INFERENCE with PROVIDED WEIGHTS.

@tjiagoM yes I did. https://github.com/ultralytics/yolov3. I achieved much much better results using their model.

this repo is good ONLY for INFERENCE with PROVIDED WEIGHTS.

No.

this repo is good ONLY for INFERENCE with PROVIDED WEIGHTS.

@Khalifa1997 maybe I was a bit in haste. One hundred images are very small sample population but can be artificially multiplied with proper data-augmentation. It depends on how many cls...

Multiple GPU

Are you using pytorch distributed package? if so, did you correctly set your default Cuda location for your local process rank? if not, this happens.

the use of self.cached in models/GCN.py

the self.cached prevented the model from recomputing the normalized graph over and over and just memorizing it since the graph is the same in inductive. Without cached, each run might...

why GPU memory cost of Mamba2 Block > full self-attention block? And how to reduce this memory cost when training?

Have you tried reducing the expansion_ratio?

Support variable-length sequences for mamba block with position indices

This PR attempts to resolve the issue derived from training with packed data?

Gradient accumulation is not efficiently implemented for distributed recipes

so in a no_sync context, gradients are accumulated in FP32?

Gradient accumulation is not efficiently implemented for distributed recipes

Extending on this (and maybe unrelated to the overall topic) In the current implementation of FSDP1, we are sharding parameters across nodes in a multi-node scenario (zero3 implementation). Is there...